2018-08-10 variable management

define inference method for get the result of forward-prob. the parameters delivered are simple
for more complicated models, the parameters should be managed by more efficient way

mechanism：construct and acquire the variables by their names

two methods

• tf.get_variable(value,shape,name)
• tf.variable_scope()

with tf.variable_scope(namespace,reuse=)
     v=tf.get_variable(name,shape)

persistence

object: tf.train.Saver()
object.save(sess,path_/model.ckpt)

generate three files,which include the graphic data and structure data
loading method

• saver=tf.train.Saver()
• saver=tf.train.import_meta_graph(path/model.ckpt/model.ckpt.meta)
  then you can obtain the graphic operation by the name of tensor like that
  sess.run(tf.get_default_graph().get_tensor_by_name(tensor_name))
• saver=tf.train.Saver([variable_name]) ,loading part of the variables
• saver=tf.train.Saver({name:newname})
  for that we can get the moving average by the dictionary mapping,like that
  saver=tf.train.Saver({'v/ExponentialMovingAverage'：v})
• saver=tf.train.Saver(ema.variable_to_restore())

presistence principles and data form

• MetaGraphDef